General Intuition

General Intuition

General Intuition builds action foundation models — systems that learn to act in dynamic environments rather than to write text. General Intuition trains on billions of action-labelled gameplay clips from Medal, the clip-sharing platform with around 17 million monthly active users, which gives it ground-truth pairs of what a player saw and what a player did. The company reports that the same model can play a Fortnite-style game for a hundred hours straight and steer a quadrupedal robot zero-shot from a single front camera after fine-tuning on eight minutes of real-world data.

Production credibility: General Intuition raised a $320M Series A at a $2.3B valuation, led by Khosla Ventures with General Catalyst, Eric Schmidt's Hillspire and Jeff Bezos participating. Reports since have placed the company in talks at a materially higher valuation, which is unconfirmed and should be treated as such. The training-data position is the concrete asset: Medal's gameplay corpus supplies action-labelled video at a scale that is difficult to reproduce, and the robot demonstration is the company's own reported result rather than an independent benchmark.

Key Features

• Action foundation models trained to operate, not to describe
• Training corpus of billions of action-labelled gameplay clips from Medal
• One model transferring between game environments and physical robots
• Reported zero-shot quadruped control from a single front-facing camera
• Fine-tuning on minutes rather than months of real-world data
• Backed by Khosla Ventures, General Catalyst, Hillspire and Jeff Bezos

Ideal Use Case

Robotics and simulation teams that need a policy able to act from vision in an unfamiliar environment, and games companies wanting agents that behave like players rather than scripted bots. As with the other action-model companies, engagement today is a commercial conversation rather than a signup, so it suits organisations evaluating partners for embodied AI work.

How General Intuition differentiates

Physical Intelligence and Skild train on robot data; World Labs and Decart build world models from video. General Intuition's bet is that gameplay is the cheapest large source of action-labelled experience — video paired with the exact inputs that produced it — and that policies learned there transfer to hardware. If that holds, the data advantage is structural rather than temporary. The counter-argument is the reality gap: games are forgiving in ways factories are not, and the transfer results published so far are the company's own.

FAQ

Q: What does General Intuition do? A: It trains action foundation models — models that learn to act in dynamic environments — using billions of action-labelled gameplay clips, and applies them to both games and robot control.

Q: How much has General Intuition raised? A: It raised a $320M Series A at a $2.3B valuation led by Khosla Ventures, with General Catalyst, Eric Schmidt's Hillspire and Jeff Bezos participating.

Q: Where does its training data come from? A: From Medal, a gameplay clip platform with around 17 million monthly active users, which supplies video paired with the player inputs that produced it.

Q: Can General Intuition's models control robots? A: The company reports that its model steered a quadrupedal robot zero-shot from a single front camera after fine-tuning on eight minutes of real-world data. That result is self-reported rather than independently benchmarked.

Q: Is there a product to sign up for? A: Not a self-serve one. General Intuition sells into robotics, simulation and games organisations through direct contact rather than published plans.

tl;dr

General Intuition trains action foundation models on billions of action-labelled gameplay clips from Medal, and reports the same model both playing games and steering a quadruped robot. It raised a $320M Series A at $2.3B led by Khosla, with Bezos and Hillspire participating.